selective_context_compressor.py

from selective_context_source import SelectiveContext
from abs_compressor import AbstractCompressor


class SCCompressor(AbstractCompressor):
    base_model = 'gpt2'

    def __init__(self, lang: str = 'en', model: str = 'gpt2', device: str = 'cpu'):
        self.sc = SelectiveContext(model_type=model, lang=lang, device=device)

    def compress(self, original_prompt: str, ratio: float = 0.7, level: str = 'phrase') -> dict:

        # count tokens of original prompt
        original_tokens = len(self.gpt_tokenizer.encode(original_prompt))

        compressed_prompt, reduced_content = self.sc(original_prompt, reduce_ratio=ratio, reduce_level=level)

        # count tokens of compressed prompt
        compressed_tokens = len(self.gpt_tokenizer.encode(compressed_prompt))

        result = {
            'compressed_prompt': compressed_prompt,
            'ratio': compressed_tokens / original_tokens,
            'original_tokens': original_tokens,
            'compressed_tokens': compressed_tokens,
            'reduced_content': reduced_content,
        }
        return result


if __name__ == '__main__':
    import time

    compressor = SCCompressor(lang='en', model='gpt2', device='cuda')
    test_prompt = "You belong to good side. In reveal phase, You can know which two players are Morgana and Assassin but you can't know which one is Morgana or Assassin specifically, you should reason it by yourself as the game progresses."

    # level can be ['sent', 'phrase', 'token']; model can be ['gpt2', 'curie'], lang can only be 'en'
    # if model is 'curie', an OPENAI-API-Key must be given in selective_context.py
    start = time.time()
    result = compressor.compress(original_prompt=test_prompt, ratio=0.4, level='phrase')
    end = time.time()
    print(result)
    print('程序运行时间为: %s Seconds' % (end - start))